Feature #1592
Fires on “print” (not labeled yet: read the examples)
- Fires on
- 0.10% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 68%
- python 26%
- story 5%
- worksheet 1%
- chat 1%
Makes the model say
s people He things ' She( It
Makes it not say
<|assistant_end|> //<|system_end|><|arg|>ablevalue /<|think_end|>
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.28py<|arg|>oldString=print(gret("Max"))<|arg|>newString=print(greet("Max
-
agent 5.18=print(next_nuumber(3))<|arg|>newString=print(next_number
-
agent 5.05oldString=print(sum_lit(SCORES))<|arg|>newString=print(sum_list
-
agent 4.94py<|arg|>oldString=print(numchars(WORD))<|arg|>newString=print(num_chars
-
agent 4.89oldString=print(is_eeven(NUMBER))<|arg|>newString=print(is_even
-
agent 4.82=print(is_eveen(SIZE))<|arg|>newString=print(is_even
-
agent 4.82<|arg|>oldString=print(filp("apple"))<|arg|>newString=print(flip("apple
-
agent 4.719, 0, 4, 5]))<|arg|>newString=print(tail([9
-
agent 4.71<|arg|>oldString=print(voels("sun"))<|arg|>newString=print(vowels("sun
-
agent 4.64oldString=print(shhout(WORD))<|arg|>newString=print(shout(WORD
-
agent 4.61maax_of(11, 2))<|arg|>newString=print(max_of
-
agent 4.59<|arg|>oldString=print(duoble(LIMIT))<|arg|>newString=print(double(LIMIT